Essays and field notes from a one-person company building systems that stay answerable for what they do. Some pieces are technical, some are structural. All of them are evidence of the thinking behind Spegling.
Founding
Storey and Wharton named the failure mode in February: cognitive surrender, intent drift, Triple Debt. Practitioners are converging on the symptoms. This essay walks the symptoms to the architecture, and lands on three primitives: Declared Intent, Living Authority, Evidence Audit.
Intelligence is becoming infrastructure. The civic layer the social contract needs at machine speed.
Part nightmare, part design document. What happens when the mirror you built learns to question the clause.
Essays varjoranta.substack.com →
Coding agents drift on design quality because nothing in training reliably rewards it. Carbon markets, geoengineering and materials discovery hit the same wall first, and none of them solved it. They built institutions around the gap instead.
A personal reflection on AI, craftsmanship, judgment, and why I still write with a pen first.
Drift is what happens when verification gets expensive. Make checking cheap enough to run constantly and behavior changes, for people and for agents alike. On why intelligence was never the bottleneck, and what a field looks like once verifying its claims gets cheap.
AI is not taking your job. It is removing the reason your job existed.
Why the best systems throw away the most. Seven forgetting strategies and the half-life of relevance.
Execution was the hard problem. Permission is the hard problem now.
The longer essay behind the thesis. What future systems must inherit.
Patterns that work. How to get agents to amplify judgment, not replace it.
Technical
The first open model at this scale trained into its 4-bit release format. Why the day-zero quant scramble is cancelled, where the trap below 4 bits sits, and what to check now that the 2.8T weights are out.
A field report from inside a vLLM quantization plugin: the three production families, the hardware moats, the MoE infrastructure gap, and the strategic inflection point one plugin is sitting on.
A follow-up to "A 70 GB model on a 48 GB MacBook." Yesterday the number was 1 tok/s. The kernel got fused. 33.8 tok/s on the same hardware class.
A field report from optimizing 3-bit quantized inference in vLLM. Three lessons on where Triton's abstractions stop paying off.
Running Qwen3.5-35B on Apple Silicon with 3-bit weight compression. What fit, what broke, and why 1 tok/s still mattered.
Field notes on 3-bit TQ3 weight compression across the Gemma 4 family.
What you actually pay for when you run inference on someone else's GPUs.
A working note on data-oblivious vector quantization, Walsh-Hadamard rotation, and where the latency goes.
Asymmetric K/V compression and progressive temporal precision for the decode-bandwidth bottleneck.
When a flat scan beats your fancy ANN index. A small lesson in scale and constants.